Friday 16 October 2015

Moore's Law at 50

Gordon Moore

Back in 1965, the country was gripped by Beatlemania, as the Fab Four released Help!, Yesterday and Nowhere Man. Across the pond in San Francisco, the nascent flower power movement was about to usher in the hippy era. NASA’s second-generation manned spacecraft, Gemini, flew for the first time, although it would be another four years before Apollo 11 landed man on the Moon. Harold Wilson was in Number 10, average UK annual income was £751, and you could buy a house for £3,660, a Mini for around £500, and a pint of bitter for 9p. Computing was dominated by million-pound mainframes and ten thousand-pound minicomputers; the microprocessor, and its promise of desktop computing, was still six years away.


While the world of 1965 is hardly recognisable today, it was in that year that Intel co-founder Gordon Moore commented on a trend in the microelectronics industry that has held true, more or less, ever since. Here we celebrate the 50th anniversary of Moore’s Law to see just what this pioneer of the electronic age foresaw, how it’s panned out, what it means to us today, and look at several other trends that have dominated the world of computer chips for the last half a century.

THE REAL MOORE’S LAW


If we’re going to commemorate Moore’s Law, we ought to make sure we get our facts right because this is one area where half truths are rife. The commonly held view that Gordon Moore predicted a doubling of processor speeds every two years, for example, doesn’t bear any resemblance to what he actually said (even though, as it so happens, this particular trend has also held true). The statement in question appeared in a 1965 article by Moore, published in the industry magazine Electronics; it suggested that “the complexity for minimum component costs has increased at a rate of roughly a factor of two per year”. This was followed by something that has the sound of a prediction about it, though Moore never referred to it as a law. “Certainly over the short term this rate can be expected to continue, if not to increase. Over the longer term, the rate of increase is a bit more uncertain, although there is no reason to believe it will not remain nearly constant for at least 10 years. That means by 1975, the number of components per integrated circuit for minimum cost will be 65,000.”

According to Moore, that figure of “complexity for minimum component costs” was 65 transistors in 1965. There were chips with fewer transistors, and there were more complicated integrated circuits in existence, but those in which the transistors worked out cheapest had 65 transistors. These components weren’t microprocessors, since no such thing existed at the time. Instead they were logic chips such as the NOR-gates, NAND-gates and inverters that were the building blocks of large-scale computers, where they were used in their thousands. A few years later the most economical chips would be solid-state memories, Intel’s first product line, and in time they would become microprocessors. So, bringing that statement into today’s terms, it wouldn’t be far off the mark to suggest that Moore’s prediction was that the number of transistors in a mainstream microprocessor would double every year.

Although we’re celebrating 50 years of his law, we shouldn’t lose sight of the fact that Moore had confidence that the trend he’d observed would hold true for only another 10 years, admitting that beyond that it was more uncertain. Indeed, in 1975, he amended his predicted rate to a doubling every two years instead of every year. Given that this was the era of the Intel 8080 and the Motorola 6800, both 8-bit microprocessors and some of the first to be used in desktop computers, that amended prediction goes back to the very birth of personal computers. As such, it gives us the opportunity to see how reality has compared to prophecy.

TRANSISTORS OF MERCY


The yellow area on the graph (left) shows how the number of transistors per microprocessor has increased since 1971 for some prominent products, including Intel’s 4004, 8080 and the x86 family. The way the yellow area represents the data isn’t too helpful because nothing seems to change until about 2001. In fact, there were huge year-on-year increases from the very start in percentage terms but, because the figures are so small in comparison to today’s billions, we just can’t see the trend at the start.

As you’ll probably recall from school maths lessons, Moore’s growth rate, defined as a doubling in a given time period, is referred to as exponential, and this sort of linear graph isn’t the best way to illustrate it. Instead, it needs to be plotted logarithmically, so the growth will appear as a straight line. The blue area is the same prominent products but now plotted on a logarithmic axis. Also plotted on the logarithmic axis is a grey line, which represents the theoretical growth rate for a doubling every two years. The closeness of the blue area – upset only by an apparent brief hiccup in the mid-90s – is testimony to Moore’s predictive abilities.

Recognising that the law that carries his name is often misquoted, and that many other aspects of microelectronics have also risen at similar rates, in a recent speech Gordon Moore joked that he was the inventor of the exponential. In reality, not everything has increased or decreased exponentially and some trends have been much more sedate. Exponential or not, though, let’s delve into some of the other astounding trends of the last few decades, trends that have either fuelled Moore’s Law or have been a consequence of it.

THE AMAZING SHRINKING CHIP


If it wasn’t for the events of the last half a century, we’d probably think that even the very first microprocessor, the Intel 4004, was a miracle of miniaturisation. With integrated circuits still in their infancy, to most electronic engineers a transistor was a little glass or metal cylinder, about 5mm in diameter and 15mm long, with three leads attached to it. Yet the 4004 was a tiny chip of silicon measuring 3x4mm, on to which no fewer than 2,300 transistors were etched.

This was achieved using a 10μm process, which means that the minimum feature size was 10 micrometres or one hundredth of a millimetre. The massive increase in the transistor count that was predicted by Gordon Moore has been accompanied by a steady decrease in the feature size, also exponential, but why was that necessary?

The first reason is fairly obvious. Today’s top-end microprocessors, such as the eight-core Core i7 Haswell-E, have 2.6 billion transistors, or just over a million times the 4004’s transistor count. Using the same feature size as the 4004, its die would measure three metres by four metres, which wouldn’t exactly lend itself to use in a server, let alone a desktop computer. This is just a start, though, as other imperatives have also been responsible for ensuring that the big squeeze remained on track.

In the realm of everyday objects, it seems that small things are invariably faster than large ones. So a humming bird is able to flap its wings much more rapidly than a golden eagle, and a short organ pipe can generate a higher pitched note, corresponding to the air vibrating more rapidly, than a long organ pipe. Much the same applies to the world of electronics. However, it’s not because the electrons have a shorter distance to travel in a smaller chip, as is sometimes suggested, since as yet this hasn’t become the overriding issue. Instead, it comes down to capacitance.

To cut a long story short, the capacitance of the transistors on a chip decreases as their dimensions decrease, and this in turn allows them to be switched on and off more quickly. And so, for many years, the size of a transistor’s gate – referred to as the process size – decreased by 30% every two years or so, so the on-chip area halved, which led to a doubling of the processor’s clock frequency.

It’s not entirely clear what the process size was back in 1965 when Moore’s Law was first postulated but, six years later, the 4004 was fabricated using a 10 micron process. Given that a human hair is about 100 microns in diameter, this sounded pretty small at the time, although that exponential decrease has made feature sizes a whole lot smaller over the years and decades. Today’s 14nm chips have linear dimensions around 700 times smaller or, in terms of their area, features have shrunk by a factor of about half a million. This resulted in an increase in clock speeds from 108kHz in 1971 to 3.8GHz in 2004 but, since clock speeds have been pretty much static ever since, it might be puzzling just why the race to produce ever smaller chips continues.

For many years, shrinking chips allowed a lower operating voltage to be used which, in turn, kept the power consumption constant even though it also permitted the clock frequency to be increased. As we’ll see when we look at operating voltages, though, this particular benefit of ever smaller chips has pretty much run out of steam. However, there’s one more benefit – one that’s been central to Moore’s Law and, while it continues today, it might just be approaching the end of the line. The very first benefit we saw of reducing the feature size was to keep the size of the chip in check, even as the number of transistors grew exponentially. Although we suggested earlier that this was due to the impracticability of using unfeasibly huge chips, economics is also a vitally important consideration. Given the huge cost of setting up a semiconductor fabrication facility, getting the maximum functionality from a chip and hence the best possible price is essential. Shrinking the size of transistors, and in so doing increasing the number you can cram on to a chip for a similar amount of work, has been a perfect way of doing exactly that.

Moore's Law

PRICES IN FREEFALL


None of the trends we’ve investigated in this article would have taken place had it not been for the laws of economics. Without year-on-year decreases in the cost of electronics’ most basic building block, the transistor, the exponential increase in their numbers that we’ve witnessed would have been unsustainable and computers would have remained the domain of the few, as some basic calculations will show.

Despite its apparently lacklustre specification and performance, the 4004 wasn’t exactly cheap when it was launched in 1971. It had a 4-bit architecture, it was clocked initially at 108kHz, rising eventually to 740kHz, it could address 32K of read-only memory but just 640 bytes of RAM. To do this, it used 2,300 transistors and cost around $200. Although 10 transistors per dollar was cheap compared to several dollars for standalone transistors, a doubling in the transistor count every couple of years would have resulted in processor prices breaking the $4,000 barrier before the end of the decade. And that doesn’t even take inflation into account. By way of contrast, today’s integrated transistors are priced in the hundred-thousandths of a penny.

The upshot of all this is that one exponentially increasing trend – that of the number of transistors per chip – has been more or less offset by an exponentially decreasing trend: the cost of those transistors. As a result, another trend – the cost of a microprocessor – really isn’t a trend at all. So, from that $200 in 1971, which probably worked out at around £1,100 here in the UK when historical exchange rates and cost of living differences are taken into account, today’s desktop processors range from around £80 to £900. Given that all our other trends have considered the most, the smallest or the fastest available on any given date, as our price for today we really ought to pick the latest and greatest, in which case prices have changed little over the whole era of the microprocessor.

If you want ever more bang for your buck, all this is good news. According to some experts, though, things are about to change. While it’s still business as usual for increases in transistor count, reductions in feature size, improvements in processor performance and so much more, some industry experts are suggesting that the cost of producing a transistor can’t fall much further. Essentially it’s getting so fiendishly difficult, and hence costly, to shrink feature sizes yet further. Indeed, it was a surprise to many when Intel confounded these analysts by reducing the cost-per-transistor in migrating to 14nm. Similar concerns are being expressed about the transition to 10nm, due in 2017, and particularly the 7nm process, for which it appears that it will be necessary, at long last, to abandon silicon technology in favour of a different semiconductor. It remains to be seen just how far this particular trend can continue.

VOLTAGE REDUCTION


Alongside all the trends we’ve seen so far, another trend has gone largely unnoticed to the lay person but has been no less important in bringing us to where we are today. That trend is a reduction in the processor’s operating voltage and, unlike the other trends, this hasn’t plummeted exponentially but much more gradually. What’s more, it’s pretty much gone as low as the laws of physics allow.

We’ve seen how shrinking the feature size keeps chip sizes to manageable proportions, even as transistor counts increase, it reduces the cost per transistor and, for many years, it allowed clock speeds to be increased. It also allows the operating voltage to be decreased and, since the electrical power consumption of a chip is related to the square of the voltage, this can bring about huge gains. In fact, for many years this effect cancelled out the increase in power consumption caused by the increased clock frequency. The relationship between feature size, clock frequency, voltage and power consumption was first examined by IBM’s Robert Dennard in 1974. It’s referred to as Dennard Scaling and can be thought of as a law that fuelled Moore’s Law for many years.

In the early days, integrated circuits such as the first logic chips and Intel’s first memory devices operated from a 5V supply. Being a very early adopter of MOS technology, the 4004 microprocessor had a 15V supply, although it wasn’t long until the teething problems were sorted out and a 5V supply became the norm for microprocessors. In fact, that 5V supply remained the de facto standard for several generations, presumably because there was no real imperative to upset the status quo, given the still low transistor count and clock speed.

In time, though, lower voltages were introduced. As 3.5V and lower voltages appeared for some of the 386s, 486s and Pentiums, and the Pentium Pro used a reduced voltage for all its variants, this trend gained momentum, but only in ever decreasing steps; a far cry from the exponential trends we see elsewhere. Today the minimum is around 0.75V, but it’s becoming increasingly difficult to quote exact figures for recent designs since different parts of the chip can operate at different voltages.

The physics might be involved but the conclusion is clear: there’s a limit to how low you can go. So, 30 years after Dennard published his paper on scaling, and 20 years after the semiconductor industry first started to embrace it, Dennard Scaling came to an end. This happened around 2004, when Intel was trying to create a 4GHz Pentium 4. Since the higher power consumption brought about by a higher clock speed could no longer be mitigated by a lower voltage, chips started to become seriously power hungry. Not only was this bad for users’ electricity bills and for the planet, but it became increasingly difficult to keep processors from frying without some serious cooling. Ever increasing clock speeds became a thing of the past, heralding the era of multicore processors.

INTO THE FUTURE


The most obvious way to draw our investigation of Moore’s Law to an end is to address the question of how much longer we can expect it to stay on track. But since this question has been raised ad infinitum whenever problems have loomed, let’s just say that the man himself recently gave it at least another 10 years, the same period he suggested half a century ago.

Perhaps a more telling conclusion would be to quote Moore in a recent interview with the Chemical Heritage Foundation. Intriguingly, rather than a prediction, he now views Moore’s Law as a self-fulfilling prophecy. “Gradually, it became something that the various industry participants recognised as the pace of innovation they had to stay on or fall behind technologically. In order to stay at the leading edge where most of the advantages of semiconductor technology get exploited, they had to move as fast as Moore’s Law predicted. So Moore’s Law went from a way of measuring what had happened to something that was kind of driving the industry.”